Residual Attention Augmented U-Shaped Network for One-Bit SAR Image Restoration

The application of one-bit sampling technology in synthetic aperture radar (SAR) systems has great potential due to its attractive advantages such as fast sampling speed, low data rate, high real-time performance, cheap hardware cost, and low energy consumption. However, one-bit sampling produces ghost targets in SAR imaging results and causes a significant reduction in the resolution and sharpness of SAR images, which is a challenge for one-bit SAR imaging. We develop a novel residual attention augmented U-shaped network (RAAUNet) with an encoder-and-decoder architecture, capable of learning the nonlinear mapping from one-bit SAR images to high-precision SAR images through end-to-end training. To enhance the efficiency of intermodule information communication at each level, our RAAUNet adopts three types of helpful skip connections that serve distinct roles in improving learning efficiency and convergence for the entire network, reducing information loss, and preserving spatial details during encoding processing, as well as transmitting multiresolution residual features. Furthermore, several specifically designed components are integrated into our network to improve its feature learning and perception abilities, where the attentive residual convolution module with the attention mechanism is employed in both encoders and decoders to endow them with the discriminative learning ability and enhance the nonlinear representation capacity, and the multiresolution fusion recovery module enriches contextual and spatial details by fusing multiresolution residual results, thereby improving the quality of the reconstructed SAR image. Numerical experiments on three synthetic one-bit SAR image datasets demonstrate that the RAAUNet achieves favorable performance against the state-of-the-art methods for one-bit SAR image restoration.